Unlocking the Potential of Confidential AI: The Synergy of AI and TEEs

Transcript

Hello perfect. Hi everyone. Thank you for being here. I know it can be hard. Some of you done some party yesterday, I think. So today I'm going to talk about confidential AI. But first let me ask you one question. Who in this audience using AI like chat GPT? I think everybody. Okay. Second question, who is using the free account account of CH GPT? Yeah, no it's not a problem. Okay, thanks. So you need to keep in mind that each time you are using free AI like chat GPT you are providing you are feeding the open AI

model and if one day something happen like a big data leak um and it happened with deepseek you will lose not lose but you will uh release all your private data and it's exactly the topic of today confidential AI so over the next 10 minutes I will talk a bit about the exec story confident confidential computing because you need to understand what is confidential computing first be before talking about confidential AI and we're going to talk to about use cases but first quick overview we started in 2017 we are OG in crypto uh we pass through two beer market then decentralized marketplace

blockchain and confidential computing and everything comes together it comes together and lead to the trust layer for deep and AI. But let's go straight to the point. Let's be a bit technical. What is confidential computing? Typical security meters protect data at rest. It mean encrypted and stored in a server or protect data in transit. It means the through when the data is traveling through different networks. But the re the real problem it's how you can protect data in use it. I mean for an application or a software the application need to see the data to use

it and it means that the data remain unencrypted in the memory and it's exactly where the malicious attack can happen and it's why we need confidential computing. Confidential computing protect data in use. It's a hardwarebased technology. Don't worry, it's the most technical slide. It's on this only one. It's a hardwarebased technology. We are using chipset, Intel chipset. And in this chipset, there is a small part of the memory able to create what we call a nclave. A nclave is basically a vault that where that everything that happen

in it remain private and confidential. Even the owner of the node, the CIS ad means the cloud admin can't see what happen in it, can alter the data. There is two key features isolation. As I said, enclave isolate the computing process and the data through physical partitioning and attestation report. It's really important because you need to prove that uh you need to verify the trustworthiness of the remote execution.

You need to prove that the data has been computed in this enclave. And regarding if we are zooming in this exec trusted execution environment, we are running two things in this enclave. Database you can store any private uh key secret API key whatever you want. It's a database running in this vault and it remain totally confidential and we can run application application able to compute any kind of data in this private environment. That's pretty cool.

We are using two technologies SJX and right now a brand new one uh I think it started in 2023 S uh TDX trust domain extension basically in this enclave in this black box with SJX you can protect an application but with TDX it's better because you can run an entire virtual machine in this enclave. So you can run whatever you want any kind of application.

So you know you understand now what is confidential computing protecting data in use. What we've done at EXC, we added the the web 3's blockchain layer and able to tokenizing data, regain ownership and add some governance rules because application running in this enclave. You can set some governance rule like who will use it, who is able to compute some data and even set a price. And the merge between confidential computing and web3 the blockchain layer is decentralized confidential computing.

But you may think like okay Martin that's cool but it sounds really complicated confidential computing. It's maybe only for hardcore developer. Not at all. We abstracted all the complexity of the complic the confidential computing and we have two tools at EXC data protector able to create some protected data. So it's a end toend encryption tool able to encrypt data and IAB generator it's a CLI common tool able to bootstrap a p any privacy preser preserving application in a minute so on the side you've got data protector able

to create some protected data encrypted data and the other side you have I generator able to create some application compatible with this TE environment and I think you get it this application are able to compute this protected data. So you keep the whole chain of trust the full chain of trust from I'm protecting my data and I compute my data with this two amazing tool. We've got some existing use case really interesting. We've got web 3 messaging.

It's a way to send email or telegram message to an Ethereum address without knowing or storing the email or or the Telegram handle. That's pretty cool. Content delivery. It's a way to transfer data and monetize data within the web3 ecosystem. Like imagine I've got my legal or health personal data. I would be able to protect my data and monetize it to someone like you want to train your AI model with my private data. you would you could do that without seeing my data thanks to our uh technology and your I app so we've got I app generator

as I said and you can bootstrap any application and then you can in one common line I app deploy you are able to deploy it in this uh confidential environment TE so let's talk a bit about AI when I started I asked a question who in the audience using uh charg openai and the free account and a lot of company are using in are using currently a free account of openi it means if an uh an employee is providing some private data from the company uh to an open AI model stuff like that you can have some

trouble and 40% of organization reported uh AI related privacy breach So it's even more important with the AI agent. We are entering in a new age. You know what is an AI assistant? But right now we are using everyday AI agent. And honestly uh I'm working with AI agent every day. What is an AI agent? It's basically AI power able to take decision on the behalf on on the on behalf of you of human. They are capable to chain action.

They are able to interact with your own tools like you can plug AI to your Gmail account to your Google agenda or whatever you want. And if you want to with this AI agent more accurate answer, more personalized experience, it means you need to provide a lot of private data. And I invite you to make a test. You can ask chat GPT like uh could you define my profile based on what you know about me and you will definitely be scared like then it know everything about you cuz you are feeling it every day and it's why it's become really critical privacy with AI it's why we are

providing confidential AI like you know now what is decentralized confidential computing it's an infrastructure And you know what is AI and the merge of both are confidential AI. That's pretty amazing because with tech TDX technology we are able to run E AV AI model AI framework too and this framework this AI model are able to compute any kind of data in this private environment and even better you can monetize every step of the AI pipeline like you can monetize your private data to train a

model or you can monetize inference of your trained model. You can monetize everything but in a private way. And we've got a pretty cool use case with Eliza OS. I think everybody here know what is Elisa OS. It's a really powerful AI framework able to do a lot of stuff because you can plug it to any kind of tool. All the community build some amazing plug-in. We are running Eliza OS in an enclave in this blackbox. So you can think about a lot of any kind of interaction on chain offchain but always in a private way things thanks to TE thanks to confidential computing

and there is a lot of use cases every day in every field like research, finance or healthcare. If you think like every day you are providing data and every day these data are definitely private. So we should be really careful right now cuz I don't know if you're I think there is a big data uh big uh breach uh two weeks ago about Google or Facebook. I think they lost uh 16 billion of uh login and password stuff like that. That's crazy. And even after bridge like that people don't realize that they need confidentiality, they need confidential computing in every field.

So I talk I talked uh uh I talked about the builder side because we are providing easy tool like it's a easy plug and play um it's package you can a builder can install it by running npm install it's for builder. So we we enable them to integrate confidential computing uh in their project. But on the end user side the cool thing it's MCP server. What is an MCP server? It's a protocol, a standard protocol uh enabling AI agent to communicate with any kind of tool like I told about like interact with Gmail, Google agenda, whatever. It thanks to this MCP server,

you can easily interact with everything without knowing um any kind of code. And with the EXC MCP server, you can use confidential computing tools, EXC tools just by prompting like could you protect my data, this data? Yeah, sure. And it will trigger data protector and protect your data and return everything about your protected data. And then you can say oh and now can you compute or provide or monetize my data with an EXC application uh created by any kind of builder you can do that with MCP server so that's pretty cool because as I said confidential computing decentralized

confidential computing can sounds quite complicated but with these all new technologies you can now interact with this technology easily just by prompting just by installing some really uh some simple package and about AI we joined the the AI unbundled alliance. We are not alone in this AI field. We are working closely with different partner like uh ITL like uh yeah it's a a big one I we are working with them that are providing computation uh computing power for for us and we are building something huge for confidential

AI. So just about what comes next in the EXC road map. I told I talk about TDX I am generator. We've got an ecosystem fund 1 million because we want to we want to um help builder grow this ecosystem build some strong confidential AI project. So if you've got a project in this field please come and contact us contact me talk to me. We will help you to build something huge with our tech. I talked about Elisa and the cool really really cool news like this privacy tech actually run on a side chain but we are deploying it on layer two it means we

are becoming the trust layer for layer that's amazing and yeah I think I have two minutes left thank you very much and if you've got any kind of question I think we yeah we've got some time saved. Thank you. Yes. Thanks uh thanks for this great talk. So in the beginning you mentioned that secrets are stored onchain on the blockchain but uh blockchain is not confidential. So how exactly do you store secrets uh on chain? Yeah that's a really cool question. Thanks. On the

data protector side, we are protect we are encrypting data. The the encrypted data is stored on IPFS or Arweave. We are working with Aweave but it's a decentralized public storage. So it's public but the data is encrypted. So we are not scared and the sciric key we used to encrypt the data we stored it in the database running in this enclave. So when an EXC application need a data to to process the data, it will get the encrypted data, it will get the symmetric key, the key uh in the database and then process the data. And about the blockchain layer, when you are protecting your data or you are deploying an EXC application, you've got

an NFT mint on the chain. So it's with the metadata, it's the the identity, it's the ownership. I talked about ownership. If I'm protecting your data, we are minting with our smart contract an NFT linked to your wallet. So you are owning this data and you can set some governance rule to set some monetization rules if you want. I hope it it's clear. Yeah. Cool. Hi, thanks for the talk. Um do you have any tools that allow onchain verification of the attestations coming out of the enclave? Okay, the

attestation uh it's a hardware based technology and we are working with Intel. So the attestation is provided by Intel. So you need for sure to trust the provider trust the provider. Uh but the attestation is provided by Intel. Yeah. So it's not onchain. Hey um do you um have you implemented a plug-in in Eliza OS to facilitate the integration of with your solution? Yeah, so we've done two two project with Eliza OS. We we are running Eliza OS in in enclave able to manage on uh Twitter account X account and we build a plug-in too. So Elisa is able to interact with

EXC tools uh and the plug-in is available in the plug-in community web but you know right now uh we are focusing on MCP server because uh Eliza OS it's a kind of MCP server at the end like you can build plug-in interact with any kind of tool and uh Eliza OS is autonomous. So yeah we've got a plugin but we are not focusing on that right now. Thank you. We will prepare for the another presentation. So the exx booth it's just uh on the right. So you can come and ask any kind of qu any question and talk about your project AI or whatever you want. See you. Thanks.